Commit 0dac827c authored by Shucai Xiao's avatar Shucai Xiao
Browse files

some preliminary optimzation for the logsoftmax function.

parent 66bae091
......@@ -76,13 +76,13 @@ device_type<T>* device_cast(T* x)
}
template <class T>
T to_hip_type(T x)
__device__ __host__ T to_hip_type(T x)
{
return x;
}
// Hip doens't support __fp16
inline float to_hip_type(gpu_half x) { return x; }
inline __device__ __host__ float to_hip_type(gpu_half x) { return x; }
} // namespace device
} // namespace gpu
......
......@@ -30,41 +30,74 @@ argument logsoftmax(hipStream_t stream,
hip_tensor_descriptor<n_dim> desc_batch(batch_shape);
hip_tensor_descriptor<n_dim> desc_data(output_shape);
// each thread is for one item in the batch
gs_launch(stream, batch_shape.elements())([=](auto i) {
auto batch_idx = desc_batch.multi(i);
auto data_idx = batch_idx;
// use one block for items in one batch.
// opt 1, load all data to lds then use the same approach as
// the current optimization
const size_t block_size = 1024;
launch(stream, batch_shape.elements() * block_size, block_size) ([=] (auto idx) __device__ {
size_t thr_idx = idx.local;
size_t blk_idx = idx.group;
// using type = typename decltype(input)::value_type;
using type = device_type<std::remove_cv_t<typename decltype(output)::value_type>>;
// get max
auto batch_max = input_ptr[desc_data.linear(batch_idx)];
for(std::size_t j = 1; j < num_in_batch; ++j)
// all data can be loaded to the lds once, so all operations are
// done in lds
MIGRAPHX_DEVICE_SHARED type lds_data[block_size + 2];
auto batch_idx = desc_batch.multi(blk_idx);
auto data_idx = batch_idx;
// load data to lds and compute the batch max
size_t item_num = num_in_batch;
lds_data[block_size] = input_ptr[0];
for (size_t i = thr_idx; i < num_in_batch; i += block_size)
{
data_idx[axis] = j;
size_t idx = desc_data.linear(data_idx);
batch_max = std::max(to_hip_type(batch_max), to_hip_type(input_ptr[idx]));
}
data_idx[axis] = i;
lds_data[i] = input_ptr[desc_data.linear(data_idx)];
for(std::size_t j = 0; j < num_in_batch; ++j)
{
data_idx[axis] = j;
size_t idx = desc_data.linear(data_idx);
output_ptr[idx] = input_ptr[idx] - batch_max;
__syncthreads();
// use thread 0 for batch_max
if (thr_idx == 0)
{
auto size = (item_num > block_size) ? block_size : item_num;
for (size_t j = 0; j < size; j++)
{
lds_data[block_size] = ::max(to_hip_type(lds_data[block_size]), to_hip_type(lds_data[j]));
}
item_num -= block_size;
}
__syncthreads();
}
auto batch_sum = ::exp(to_hip_type(output_ptr[desc_data.linear(batch_idx)]));
for(std::size_t j = 1; j < num_in_batch; ++j)
const size_t block_size1 = block_size + 1;
lds_data[block_size1] = 0;
item_num = num_in_batch;
for (size_t i = thr_idx; i < num_in_batch; i += block_size)
{
data_idx[axis] = j;
size_t idx = desc_data.linear(data_idx);
batch_sum += ::exp(to_hip_type(output_ptr[idx]));
data_idx[axis] = i;
lds_data[i] = input_ptr[desc_data.linear(data_idx)];
__syncthreads();
// use thread 0 for batch_max
if (thr_idx == 0)
{
auto size = (item_num > block_size) ? block_size : item_num;
for (size_t j = 0; j < size; j++)
{
lds_data[block_size1] += ::exp(to_hip_type(lds_data[j] - lds_data[block_size]));
}
item_num -= block_size;
}
__syncthreads();
}
batch_sum = ::log(to_hip_type(batch_sum));
for(std::size_t j = 0; j < num_in_batch; ++j)
auto log_batch_sum = ::log(to_hip_type(lds_data[block_size1])) + lds_data[block_size];
item_num = num_in_batch;
for (size_t i = thr_idx; i < num_in_batch; i += block_size)
{
data_idx[axis] = j;
size_t idx = desc_data.linear(data_idx);
output_ptr[idx] -= batch_sum;
data_idx[axis] = i;
size_t index = desc_data.linear(data_idx);
output_ptr[index] = input_ptr[index] - log_batch_sum;
}
});
});
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment